Hi everyone. In my research I am willing to understand the moderating impact of one variable. However, I am only interested in the interaction effect and not in the main effect. If you can please help. For e.g.
regress y x1 x2 x2*x3 is this allowed?
or is it mandatory to consider the main effects too like the following:
regress y x1 x2 x3 x2*x3
Thank you.
@Supartim First condition for testing the interaction/moderation effect is that there is an existing relationship between X and Y variables. Interaction effect is nothing but suppression/enhancement effect on the direct relationship between X and Y and without taking direct effect into consideration how would you be draw inference for X>Y relationship at different levels of the moderator?
Thanks for sharing your answers. Let me clarify the problem a bit more. The research requires to find out the impact of a moderated moderation on a relationship. Hence the equation is like the following:
regress y x1 x2 x3 x4 x2*x3*x4.
The problem with the above approach is it considers all the possible combinations like x2*x3, x3*x4, x2*x4, x2*x3*x4. This is also distorting the outcome.
The theory seems to suggest only x2*x3 and x2*x3*x4 to be valid considerations. Can we consider only these two interactions? Or is it mandatory to consider all other combinations (which comes out by default).
In STATA we can select specific combinations using "#" between the variables. Whereas the default "##" will show all the possible combinations.
Thank you.
This really is a subject-matter question. If the theory proposes a particular model, then this is what you should use. Your simpler model would give you an estimate for x2*x3*x4 that is unadjsuted for any possible x2*x4 and x3*x4 interaction. This may or may not make sense. But this is not a statistical question. Knowing nothing about the underlying theory, David's advice is good (using the "full model" including all possible lowe-order interactions), as it' is usually saver to interpret the fully adjusted higher-oder interaction.
You may get an impression when comparing a simple linear regression with and without intercept. Both can make sense, and which is appropriate is not a statistical question but depends on the problem and subject theory. The slope is always interpretable in view of the data if the intercept is included, but there can be occasions when it's more appropriate to have no intercept. Depending on the actual data it is possible that the line is a bad fit to the data - but if I know that the intercept must be zero, it also displays the problems in the data (as may be seen in residual diagnostic plots).
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression